Memreflex: Adaptive Flashcards for Mobile Microlearning
Total Page：16
File Type：pdf, Size：1020Kb
MemReflex: Adaptive Flashcards for Mobile Microlearning Darren Edge1 Stephen Fitchett1,2 Michael Whitney1,3 James Landay1,4 1 Microsoft Research 2 University of 3 University of North 4 University of Asia Canterbury Carolina Charlotte Washington Beijing, China New Zealand USA USA [email protected], [email protected], [email protected], [email protected] ABSTRACT Flashcard systems typically help students learn facts (e.g., definitions, names, and dates), relying on intense initial memoriztion with subsequent tests delayed up to days later. This approach does not exploit the short, sparse, and mobile opportunities for microlearning throughout the day, nor does it support learners who need the motivation that comes from successful study sessions. In contrast, our MemReflex system of adaptive flashcards gives fast-feedback by retesting new items in quick succession, dynamically scheduling future tests according to a model of the learner’s memory. We evaluate MemReflex across three user studies. Figure 1. Adaptive flashcards. A cue (left) triggers recall of In the first two studies, we demonstrate its effectiveness for target information (right). Adaptive scheduling of cued recall both audio and text modalities, even while walking and tests raises response accuracies towards a goal level, e.g., 90%. distracted. In the third study of second-language vocabulary learning, we show how MemReflex enhanced learner We believe mobile learning should be context-aware in a accuracy, confidence, and perceptions of control and broad sense – sensitive to learner history as well as the success. Overall, the work suggests new directions for immediate cues of time, location, and motion. We also mobile microlearning and “micro activities” in general. believe that both text and audio interaction modalities are important to engage learners with different learning styles, Author Keywords as well as learners who move between contexts where Mobile Flashcards; Adaptive Systems; Language Learning different modalities are most appropriate (for example, it ACM Classification Keywords might be safer to listen via headphones when navigating H.5 Information interfaces and presentation: User Interfaces busy public spaces, but politer to read from the screen in social situations where some conversation is anticipated). General Terms Algorithms; Design; Experimentation; Human Factors This paper presents a systematic investigation of how INTRODUCTION learner motivation for microlearning using mobile The mobile phone is the ideal platform for long-term flashcards is affected by adaptation to past performances, as learning, being portable, individual, unobtrusive, available, well as how learner performance “on the move” is affected adaptable, persistent, and useful [22]. In particular, mobile by the selection of interaction modality. The contribution to phones can support microlearning [13] in fragments of free mobile HCI is a demonstration of how adaptive flashcards time throughout the day. Previous work in HCI has can support text and audio-based mobile learning even examined how flashcards can support mobile microlearning when walking, and how such adaptation helps drive learner of second-language phrases presented in context, motivation both in the moment and over the longer-term. investigating what material should be studied where [12]. We begin with a literature review that motivates system However, relatively little attention has been paid to when features, before illustrating how existing flashcard systems items should be introduced and reviewed based on how the do not account for the special characteristics of mobile learner has performed in past microlearning sessions, microlearning. Next, we present the algorithm and especially in terms of how this relates to learner motivation. interface design of our adaptive flashcard system, which we call MemReflex. These flashcards present cues from which learners attempt to recall the target information, with tests Permission to make digital or hard copies of all or part of this work for scheduled according to a model of the learner’s memory personal or classroom use is granted without fee provided that copies are (see Figure 1). We end with three user studies that show the not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, effectiveness of MemReflex from the immediate to the or republish, to post on servers or to redistribute to lists, requires prior longer-term, using either audio or text modalities, even specific permission and/or a fee. while distracted by the demands of learning on the move. MobileHCI’12, September 21–24, 2012, San Francisco, CA, USA. Copyright 2012 ACM 978-1-4503-1105-2/12/09...$10.00. RELATED WORK In this section, we survey both the findings of learning research and the learning systems that aim to exploit them. Although we focus on approaches to learning that can be exploited in mobile contexts, our contribution itself is to the larger body of work on mobile learning or m-learning1. Learning Research Learning is not a single process, but a hierarchy of processes reflecting progressive orders of change [2]. Zero order learning is characterized by responses without correction; first order learning by correction of errors within Figure 2. Adaptive spaced-repetition learning sets of alternatives; and second order learning by a change in the sets of alternatives or the distribution of first order The forgetting curve learning over time (also known as “learning to learn”). The forgetting curve, discovered in 1885 [11], describes the inverse exponential nature of forgetting. In the presence of Reviewing a physical flashcard is an example of first order repeated, spaced repetitions as described above, the learning: the front side of the card acts as a cue for the strength of a memory is increased, resulting in a more target on the reverse. When a learner attempts to recall the gradual process of forgetting. The most sophisticated target given such a cue, the learning style is called cued psychological modeling of this process is derived from the recall and has been studied extensively in the learning ACT-R activation-based model of declarative memory, literature. In contrast, our investigation of how to motivate with each test of an item introducing a new memory trace learners to appropriate time for microlearning is a question whose decay rate is a power law function of all traces for of second order learning that has yet to be fully explored. the item at the time of the test [17]. In this model, the The testing effect higher an item’s activation, the smaller the effect any Much research into learning investigates and exploits the additional tests will have on its long-term retention. testing effect – that tests strengthen memory more than Overlearning extra opportunities to study, even when mental retrieval is Once a learner correctly recalls an item using cued recall, not accompanied by an outward response. Such test- any further testing of the same item in the same session is directed learning can therefore take place in contexts where described as overlearning. In two experiments and a review it is undesirable to produce overt responses, such as in of the overlearning literature, Rohrer et al. [20] show that public places. Moreover, it has been demonstrated across a within a single learning session, doubling the number of variety of domains, including the learning of vocabulary in tests for each item typically more than doubles the native and second languages, face–name associations, percentage of correct responses when tested again one week general facts, text passages, word lists, and even maps [6]. later. They suggest that such overlearning is necessary for Cued recall tests of the form A? have also been shown to short-term retention in situations such as preparing for an enhance retention in the opposite direction B?, as well as exam later in the day or learning foreign language enhancing free recall of all cues (As) and targets (Bs) [7]. vocabulary in advance of planned conversations. However, The spacing effect these benefits diminish for longer retention intervals. The spacing effect is that when learning a set of items, Learning Systems superior retention results from multiple shorter Since 1967, the dominant approach to audio language presentations than from a single “massed” presentation. The learning has been the Pimsleur System [19]. This is a series time separating different study episodes of the same of 30-minute audio lessons in which basic vocabulary and material is known as the inter-study interval or ISI. Studies phrases are introduced and reviewed in cued-recall fashion of the spacing effect typically manipulate the ISI of two according to a schedule of graduated interval recall. This is study episodes, and compare their effect on a later test that a progressive series of exponentially expanding intervals occurs after a fixed retention interval. In a review of 427 (repetitions after 51s = 5 seconds, 52s = 25 seconds, 53s 2 articles on cued recall learning, it was found that the minutes, 54s 10 minutes) both within and (roughly) across optimal ISI increases as the retention interval increases [8]. lessons. The advantages are that it uses deliberate For example, the optimum ISI for a 1-minute retention overlearning to breed confidence and support immediate interval was less than 1 minute, whereas for retention language use. The disadvantages of such lessons, however, intervals of 6 months or more it was at least 1 month. The are that repetitions are not scheduled in real-time, there is implication is that multiple study episodes are needed for no adaptation to user feedback, lesson-length chunks of free continuous retention, as shown in Figure 2. time must be available, and ultimately the learner will run out of lessons. This latter difficulty is addressed by Gradint [14], which allows learners to make their own Pimsleur-like 1 See the m-learning website at http://www.m-learning.org/.